![]() | Only 14 pages are availabe for public view |
Abstract Image segmentation has become a crucial part in many computer vision applications. Differ- ent algorithms have been proposed for automatically segmenting images into homogeneous regions. However, fully automatic solutions are not able to extract semantic objects directly. Therefore, there has been an extensive research effort on a class of segmentation algorithms that takes human interaction into account in order to enhance the accuracy and usability of segmentation algorithms. In this thesis, the drawbacks of a recently proposed interactive image segmentation algo- rithm are addressed. The algorithm considered is the Maximal Similarity Region Merging (MSRM) algorithm. Although this algorithm was found to provide very good results in terms of accuracy, it was found to suffer from major drawbacks that much affect its time perfor- mance in many cases and accuracy in some cases. These drawbacks are mainly due to its dependence on a fixed initial automatic segmentation that may not be suitable enough for the object being extracted, and also due to inefficient region merging procedure that results into much execution time. To avoid the drawbacks of the MSRM algorithm, an adaptive initial segmentation scheme , was developed using variable-bandwidth mean shift image segmentation in order to provide a suitable initial segmentation that guarantees both accuracy and speed. In addition, a more efficient region merging scheme was devised to reduce the execution time of the algorithm. Moreover, a parallel version of the algorithm was proposed in order to enhance the speed of the algorithm. The experimental evaluation conducted over more than 90 different images shows significant improvement over the original MSRM segmentation algorithm in terms of speed and accuracy. |